One Line at a Time — Generation and Internal Evaluation of Interactive Poetry

We present methods that produce poetry one line at a time, in a manner that allows simple interaction in human-computer co-creative poetry writing. The methods are based on fine-tuning sequence-to-sequence neural models, in our case mBART. We also consider several internal evaluation measures by which an interactive system can assess and filter the lines it suggests to the user. These measures concern the coherence, tautology, and diversity of the candidate lines. We empirically validate two of them and apply three on the mBART-based poetry generation methods. The results suggest that fine-tuning a pre- trained sequence-tosequence model is a feasible approach, and that the internal evaluation measures help select suitable models as well as suitable lines.